Core Concepts
Enhancing manifold regularization through label propagation and Neumann heat kernel for improved classification performance.
Abstract
The content introduces a semi-supervised learning model based on manifold regularization, focusing on improving it through label propagation and the Neumann heat kernel. It discusses the shortcomings of the original model, proposes enhancements, and validates them through experiments. The paper outlines the construction of a label propagation model, explains the diffusion mapping algorithm, and presents numerical experiments to compare different models' performance in various datasets.
Introduction
Overview of Semi-Supervised Learning Algorithms.
Manifold Regularization Model
Loss function components and geometric features.
Label Propagation Model
Construction process and diffusion process explanation.
Neumann Heat Kernel Regularized Least Squares (NHKRLS)
Introduction to NHKRLS model and its optimization.
Numerical Experiments
Performance evaluation on generated datasets and MNIST dataset.
Binary Classification Task
Comparison between NHKRLS, LapRLS, and LS models.
Multi-Classification Task
Evaluation of NHKRLS performance in multi-classification tasks.
Stats
"We use the Frobenius norm to construct the distance matrix between different data points."
"For the NHKRLS algorithm, we chose 10 diffusion steps considering only the 5-nearest neighbors in each diffusion."
Quotes
"The NHKRLS model consistently demonstrates higher classification accuracy."
"The LapRLS model shows lower classification accuracy when the number of labeled samples is too low."